Tokens and Meaning

Last weekend I burned through a prototype in a single sitting. An idea I’d been kicking around for a while, a small tool, nothing ambitious. I pointed Claude Code at it, described what I wanted, and watched it materialize. Two hours, maybe three. It worked. It ran. It even looked decent.

By Sunday night I was staring at it with the quiet recognition that I’d built something nobody asked for and nobody would use.

I’d gone fast to nowhere.

The Speed Illusion

MIT Technology Review named generative coding one of the breakthrough technologies of 2026. Boris Cherny, who leads Claude Code at Anthropic, told Lenny Rachitsky that coding is largely solved. Claude Code alone now accounts for 4% of public GitHub commits. Four percent. Of all public commits. From a single tool.

The implementation bottleneck that defined our careers is dissolving. What opens up on the other side is supposed to be judgment, taste, architectural thinking: the things engineers always said they wanted more time for. In theory, that sounds like liberation.

In practice, it feels more like vertigo. When the thing you were good at becomes cheap, you have to reckon with what’s left. What’s left is the question we’ve always been bad at answering: should we build this at all?

That question turns out to be much harder than the code.

Shipping Numbers

Karri Saarinen, the CEO of Linear, put it bluntly: we’ve lost our sense of judgment and moderation when it comes to product building. Token burn rate, prototype count, percentage of agent-written code. These have become the new vanity metrics. The moment you turn them into celebrated measures of progress, you stop shipping value and start shipping numbers.

My weekend prototype wasn’t really about solving a problem. It was about the thrill of making something appear. LLM-generated prototypes feel like late-night whiteboarding sessions: exciting in the moment, impressive in their speed, and by Wednesday you realize the idea was shallow, distracting, or simply wrong.

More power to build should not reduce our need to think. It should increase it. Not everything needs to be agent-shaped. A scheduled task does not need a full LLM sandbox. Making something agentic because it feels current doesn’t make it right-sized or effective.

Linear didn’t become a billion-dollar company by shipping everything they could. They became one by shipping with restraint, focusing on craft, and saying no to most things. That’s harder to celebrate on a dashboard.

Launching a product is trivial in 2026. Standing out is not.

The Real Bottleneck

If coding is solved, what’s the bottleneck? I keep coming back to product thinking. Understanding markets, customers, competitive dynamics. Knowing where to point the machine before you turn it on.

I’ve been reading Swipe to Unlock by Neel Mehta, Parth Detroja, and Aditya Agashe. A primer on technology and business strategy written by PMs at Google, Microsoft, and Facebook. The book predates AI-native development, but its core premise lands harder now: you need to understand the technology landscape to make good product decisions. Platform thinking, network effects, data as a strategic asset. These aren’t engineering problems. They’re product problems. AI can’t shortcut them for you.

Cherny frames the shift as going from “engineer” to “builder.” I like that framing. But it carries a weight people gloss over. Builder implies you know what you’re building and for whom. Technical fluency still matters; it informs strategy, it helps you see what’s possible. But fluency without direction is just speed.

And I’ve already seen what speed without direction looks like.

What I Don’t Have

I’ve spent years getting better at how to build things. Cleaner code, better architecture, faster iteration. The gap in my toolkit isn’t technical. It’s knowing what to build and why.

I don’t have a framework to offer. What I will say is that I’ve started reading more about markets and customers than about frameworks and languages. I’ve started asking “who is this for?” before I ask “how do I build it?” Not because someone told me to, but because an empty prototype on a Sunday night has a way of making things obvious.

The tokens are cheap. Meaning is not.